结合天空识别和暗通道原理的图像去雾

Image haze removal based on sky region detection and dark channel prior

目的目前较为流行的去雾算法对天空区域的处理效果不佳,容易造成方块效应以及色彩严重失真。针对该问题,提出一种基于天空识别和暗通道原理的单幅图像去雾方法。方法天空识别算法,将雾图分为天空与非天空部分,对其分别估计透射率图,通过大气散射模型得到复原图像;针对利用暗通道原理得到的去雾图像普遍偏暗的问题,对其进行色彩重映射,以增加图像亮度,提升图像视觉效果。结果大量实验结果表明,本文算法复原的图像清晰自然,尤其是天空区域平滑明亮,取得了很好的去雾效果。结论基于天空识别,提出了一种新颖的单幅图像去雾算法。与He Kaiming以及Tarel的算法相比,去雾后图像整体效果更佳。

Objective Images captured in foggy weather are of ten degraded byatmospheric absorption and scattering. Haze removal is highly desired in image processing and computer vision applications. Removing haze can significantly increase the visibility of the scene. In addition, most image processing and computer vision algorithms usually assume that the input image is the scene radiance. Therefore, several methods for haze removal have been proposed. However, the sky region processed by most of these algorithms is degraded by block noise and serious color distortion. To address this issue, this pa- per proposes an improved single image haze removal method based on sky region detection and dark channel prior. Method Our proposed method consists of three major stages： sky region detection, haze removal, and tone mapping. In the first stage, sky is usually a large and smooth region with high intensity. On the basis of these characteristics, an effective algo- rithm is designed to divide the input image into ＂sky＂ and ＂non-sky＂ regions. In the next stage, dark channel prior is used to estimate the transmission maps of the two regions, and a guided filter is applied to refine these maps, such that the haze-free image can be recovered by the atmosphere scattering model. The final stage uses a simple tone mapping algorithm to increase the image brightness, which leads to good visual effects. Result In He＇s haze removal algorithm, dark channel prior is no longer a good prior because sky regions may have high intensity. Consequently, the sky region of the recovered haze-free image will have serious noise and color distortion. We combine sky region detection and dark channel prior to eliminate noise and color distortion. Several experiments show that images restored by the proposed algorithm are clear and natural. In particular, the sky region is smooth and bright. Conclusion Dark channel prior is a very good prior for image haze removal, but it is not suitable for sky regions because it leads to block noise and serious color distortion. On the basis of sky region detection, a novel single image haze removal algorithm is proposed in this paper. The presented algorithm can achieve better results than the defogging algorithms proposed by He Kaiming and Tarel.